2 * Copyright (C) 1994, David Greenman
3 * Copyright (c) 1990, 1993
4 * The Regents of the University of California. All rights reserved.
6 * This code is derived from software contributed to Berkeley by
7 * the University of Utah, and William Jolitz.
9 * Redistribution and use in source and binary forms, with or without
10 * modification, are permitted provided that the following conditions
12 * 1. Redistributions of source code must retain the above copyright
13 * notice, this list of conditions and the following disclaimer.
14 * 2. Redistributions in binary form must reproduce the above copyright
15 * notice, this list of conditions and the following disclaimer in the
16 * documentation and/or other materials provided with the distribution.
17 * 3. All advertising materials mentioning features or use of this software
18 * must display the following acknowledgement:
19 * This product includes software developed by the University of
20 * California, Berkeley and its contributors.
21 * 4. Neither the name of the University nor the names of its contributors
22 * may be used to endorse or promote products derived from this software
23 * without specific prior written permission.
25 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
26 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
27 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
28 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
29 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
30 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
31 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
32 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
33 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
34 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
37 * from: @(#)trap.c 7.4 (Berkeley) 5/13/91
38 * $FreeBSD: src/sys/i386/i386/trap.c,v 1.147.2.11 2003/02/27 19:09:59 luoqi Exp $
39 * $DragonFly: src/sys/i386/i386/Attic/trap.c,v 1.64 2005/10/11 09:59:56 corecode Exp $
43 * 386 Trap and System call handling
51 #include "opt_ktrace.h"
52 #include "opt_clock.h"
55 #include <sys/param.h>
56 #include <sys/systm.h>
58 #include <sys/pioctl.h>
59 #include <sys/kernel.h>
60 #include <sys/resourcevar.h>
61 #include <sys/signalvar.h>
62 #include <sys/syscall.h>
63 #include <sys/sysctl.h>
64 #include <sys/sysent.h>
66 #include <sys/vmmeter.h>
67 #include <sys/malloc.h>
69 #include <sys/ktrace.h>
71 #include <sys/upcall.h>
72 #include <sys/sysproto.h>
73 #include <sys/sysunion.h>
76 #include <vm/vm_param.h>
79 #include <vm/vm_kern.h>
80 #include <vm/vm_map.h>
81 #include <vm/vm_page.h>
82 #include <vm/vm_extern.h>
84 #include <machine/cpu.h>
85 #include <machine/ipl.h>
86 #include <machine/md_var.h>
87 #include <machine/pcb.h>
89 #include <machine/smp.h>
91 #include <machine/tss.h>
92 #include <machine/globaldata.h>
94 #include <i386/isa/intr_machdep.h>
97 #include <sys/syslog.h>
98 #include <machine/clock.h>
101 #include <machine/vm86.h>
104 #include <sys/msgport2.h>
105 #include <sys/thread2.h>
107 int (*pmath_emulate) (struct trapframe *);
109 extern void trap (struct trapframe frame);
110 extern int trapwrite (unsigned addr);
111 extern void syscall2 (struct trapframe frame);
112 extern void sendsys2 (struct trapframe frame);
113 extern void waitsys2 (struct trapframe frame);
115 static int trap_pfault (struct trapframe *, int, vm_offset_t);
116 static void trap_fatal (struct trapframe *, vm_offset_t);
117 void dblfault_handler (void);
119 extern inthand_t IDTVEC(syscall);
121 #define MAX_TRAP_MSG 28
122 static char *trap_msg[] = {
124 "privileged instruction fault", /* 1 T_PRIVINFLT */
126 "breakpoint instruction fault", /* 3 T_BPTFLT */
129 "arithmetic trap", /* 6 T_ARITHTRAP */
130 "system forced exception", /* 7 T_ASTFLT */
132 "general protection fault", /* 9 T_PROTFLT */
133 "trace trap", /* 10 T_TRCTRAP */
135 "page fault", /* 12 T_PAGEFLT */
137 "alignment fault", /* 14 T_ALIGNFLT */
141 "integer divide fault", /* 18 T_DIVIDE */
142 "non-maskable interrupt trap", /* 19 T_NMI */
143 "overflow trap", /* 20 T_OFLOW */
144 "FPU bounds check fault", /* 21 T_BOUND */
145 "FPU device not available", /* 22 T_DNA */
146 "double fault", /* 23 T_DOUBLEFLT */
147 "FPU operand fetch fault", /* 24 T_FPOPFLT */
148 "invalid TSS fault", /* 25 T_TSSFLT */
149 "segment not present fault", /* 26 T_SEGNPFLT */
150 "stack fault", /* 27 T_STKFLT */
151 "machine check trap", /* 28 T_MCHK */
154 #if defined(I586_CPU) && !defined(NO_F00F_HACK)
155 extern int has_f00f_bug;
159 static int ddb_on_nmi = 1;
160 SYSCTL_INT(_machdep, OID_AUTO, ddb_on_nmi, CTLFLAG_RW,
161 &ddb_on_nmi, 0, "Go to DDB on NMI");
163 static int panic_on_nmi = 1;
164 SYSCTL_INT(_machdep, OID_AUTO, panic_on_nmi, CTLFLAG_RW,
165 &panic_on_nmi, 0, "Panic on NMI");
166 static int fast_release;
167 SYSCTL_INT(_machdep, OID_AUTO, fast_release, CTLFLAG_RW,
168 &fast_release, 0, "Passive Release was optimal");
169 static int slow_release;
170 SYSCTL_INT(_machdep, OID_AUTO, slow_release, CTLFLAG_RW,
171 &slow_release, 0, "Passive Release was nonoptimal");
173 MALLOC_DEFINE(M_SYSMSG, "sysmsg", "sysmsg structure");
174 extern int max_sysmsg;
177 * Passive USER->KERNEL transition. This only occurs if we block in the
178 * kernel while still holding our userland priority. We have to fixup our
179 * priority in order to avoid potential deadlocks before we allow the system
180 * to switch us to another thread.
183 passive_release(struct thread *td)
185 struct lwp *lp = td->td_lwp;
187 td->td_release = NULL;
188 lwkt_setpri_self(TDPRI_KERN_USER);
189 lp->lwp_proc->p_usched->release_curproc(lp);
193 * userenter() passively intercepts the thread switch function to increase
194 * the thread priority from a user priority to a kernel priority, reducing
195 * syscall and trap overhead for the case where no switch occurs.
199 userenter(struct thread *curtd)
201 curtd->td_release = passive_release;
205 * Handle signals, upcalls, profiling, and other AST's and/or tasks that
206 * must be completed before we can return to or try to return to userland.
208 * Note that td_sticks is a 64 bit quantity, but there's no point doing 64
209 * arithmatic on the delta calculation so the absolute tick values are
210 * truncated to an integer.
213 userret(struct lwp *lp, struct trapframe *frame, int sticks)
215 struct proc *p = lp->lwp_proc;
219 * Post any pending upcalls
221 if (p->p_flag & P_UPCALLPEND) {
222 p->p_flag &= ~P_UPCALLPEND;
227 * Post any pending signals
229 while ((sig = CURSIG(p)) != 0) {
234 * Charge system time if profiling. Note: times are in microseconds.
236 if (p->p_flag & P_PROFIL) {
237 addupc_task(p, frame->tf_eip,
238 (u_int)((int)p->p_thread->td_sticks - sticks));
242 * Post any pending signals XXX
244 while ((sig = CURSIG(p)) != 0)
249 * Cleanup from userenter and any passive release that might have occured.
250 * We must reclaim the current-process designation before we can return
251 * to usermode. We also handle both LWKT and USER reschedule requests.
254 userexit(struct lwp *lp)
256 struct thread *td = lp->lwp_thread;
257 globaldata_t gd = td->td_gd;
261 * If a user reschedule is requested force a new process to be
262 * chosen by releasing the current process. Our process will only
263 * be chosen again if it has a considerably better priority.
265 if (user_resched_wanted())
266 lp->lwp_proc->p_usched->release_curproc(lp);
271 * Handle a LWKT reschedule request first. Since our passive release
272 * is still in place we do not have to do anything special.
274 if (lwkt_resched_wanted())
278 * Acquire the current process designation if we do not own it.
279 * Note that acquire_curproc() does not reset the user reschedule
280 * bit on purpose, because we may need to accumulate over several
281 * threads waking up at the same time.
283 * NOTE: userland scheduler cruft: because processes are removed
284 * from the userland scheduler's queue we run through loops to try
285 * to figure out which is the best of [ existing, waking-up ]
288 if (lp != gd->gd_uschedcp) {
290 lp->lwp_proc->p_usched->acquire_curproc(lp);
291 /* We may have switched cpus on acquisition */
298 * Reduce our priority in preparation for a return to userland. If
299 * our passive release function was still in place, our priority was
300 * never raised and does not need to be reduced.
302 if (td->td_release == NULL)
303 lwkt_setpri_self(TDPRI_USER_NORM);
304 td->td_release = NULL;
307 * After reducing our priority there might be other kernel-level
308 * LWKTs that now have a greater priority. Run them as necessary.
309 * We don't have to worry about losing cpu to userland because
310 * we still control the current-process designation and we no longer
311 * have a passive release function installed.
313 if (lwkt_checkpri_self())
317 * If a userland reschedule is [still] pending we may not be the best
318 * selected process. Select a better one. If another LWKT resched
319 * is pending the trap will be re-entered.
321 if (user_resched_wanted()) {
322 lp->lwp_proc->p_usched->select_curproc(gd);
323 if (lp != gd->gd_uschedcp) {
324 lwkt_setpri_self(TDPRI_KERN_USER);
331 * Exception, fault, and trap interface to the kernel.
332 * This common code is called from assembly language IDT gate entry
333 * routines that prepare a suitable stack frame, and restore this
334 * frame after the exception has been processed.
336 * This function is also called from doreti in an interlock to handle ASTs.
337 * For example: hardwareint->INTROUTINE->(set ast)->doreti->trap
339 * NOTE! We have to retrieve the fault address prior to obtaining the
340 * MP lock because get_mplock() may switch out. YYY cr2 really ought
341 * to be retrieved by the assembly code, not here.
343 * XXX gd_trap_nesting_level currently prevents lwkt_switch() from panicing
344 * if an attempt is made to switch from a fast interrupt or IPI. This is
345 * necessary to properly take fatal kernel traps on SMP machines if
346 * get_mplock() has to block.
350 struct trapframe frame;
352 struct globaldata *gd = mycpu;
353 struct thread *td = gd->gd_curthread;
354 struct lwp *lp = td->td_lwp;
357 int i = 0, ucode = 0, type, code;
363 eva = (frame.tf_trapno == T_PAGEFLT ? rcr2() : 0);
364 ++gd->gd_trap_nesting_level;
366 trap_fatal(&frame, eva);
367 --gd->gd_trap_nesting_level;
373 ++gd->gd_trap_nesting_level;
374 if (frame.tf_trapno == T_PAGEFLT) {
376 * For some Cyrix CPUs, %cr2 is clobbered by interrupts.
377 * This problem is worked around by using an interrupt
378 * gate for the pagefault handler. We are finally ready
379 * to read %cr2 and then must reenable interrupts.
381 * XXX this should be in the switch statement, but the
382 * NO_FOOF_HACK and VM86 goto and ifdefs obfuscate the
383 * flow of control too much for this to be obviously
392 --gd->gd_trap_nesting_level;
394 * MP lock is held at this point
397 if (!(frame.tf_eflags & PSL_I)) {
399 * Buggy application or kernel code has disabled interrupts
400 * and then trapped. Enabling interrupts now is wrong, but
401 * it is better than running with interrupts disabled until
402 * they are accidentally enabled later.
404 type = frame.tf_trapno;
405 if (ISPL(frame.tf_cs)==SEL_UPL || (frame.tf_eflags & PSL_VM)) {
407 "pid %ld (%s): trap %d with interrupts disabled\n",
408 (long)curproc->p_pid, curproc->p_comm, type);
409 } else if (type != T_BPTFLT && type != T_TRCTRAP) {
411 * XXX not quite right, since this may be for a
412 * multiple fault in user mode.
414 printf("kernel trap %d with interrupts disabled\n",
420 #if defined(I586_CPU) && !defined(NO_F00F_HACK)
423 type = frame.tf_trapno;
427 if (frame.tf_eflags & PSL_VM &&
428 (type == T_PROTFLT || type == T_STKFLT)) {
430 KKASSERT(td->td_mpcount > 0);
432 i = vm86_emulate((struct vm86frame *)&frame);
434 KKASSERT(td->td_mpcount > 0);
438 * returns to original process
440 vm86_trap((struct vm86frame *)&frame);
447 * these traps want either a process context, or
448 * assume a normal userspace trap.
452 trap_fatal(&frame, eva);
455 type = T_BPTFLT; /* kernel breakpoint */
458 goto kernel_trap; /* normal kernel trap handling */
461 if ((ISPL(frame.tf_cs) == SEL_UPL) || (frame.tf_eflags & PSL_VM)) {
466 sticks = (int)td->td_sticks;
467 lp->lwp_md.md_regs = &frame;
470 case T_PRIVINFLT: /* privileged instruction fault */
475 case T_BPTFLT: /* bpt instruction fault */
476 case T_TRCTRAP: /* trace trap */
477 frame.tf_eflags &= ~PSL_T;
481 case T_ARITHTRAP: /* arithmetic trap */
486 case T_ASTFLT: /* Allow process switch */
487 mycpu->gd_cnt.v_soft++;
488 if (mycpu->gd_reqflags & RQF_AST_OWEUPC) {
489 atomic_clear_int_nonlocked(&mycpu->gd_reqflags,
491 addupc_task(p, p->p_prof.pr_addr,
497 * The following two traps can happen in
498 * vm86 mode, and, if so, we want to handle
501 case T_PROTFLT: /* general protection fault */
502 case T_STKFLT: /* stack fault */
503 if (frame.tf_eflags & PSL_VM) {
504 i = vm86_emulate((struct vm86frame *)&frame);
511 case T_SEGNPFLT: /* segment not present fault */
512 case T_TSSFLT: /* invalid TSS fault */
513 case T_DOUBLEFLT: /* double fault */
515 ucode = code + BUS_SEGM_FAULT ;
519 case T_PAGEFLT: /* page fault */
520 i = trap_pfault(&frame, TRUE, eva);
523 #if defined(I586_CPU) && !defined(NO_F00F_HACK)
533 case T_DIVIDE: /* integer divide fault */
541 goto handle_powerfail;
542 #else /* !POWERFAIL_NMI */
543 /* machine/parity/power fail/"kitchen sink" faults */
544 if (isa_nmi(code) == 0) {
547 * NMI can be hooked up to a pushbutton
551 printf ("NMI ... going to debugger\n");
552 kdb_trap (type, 0, &frame);
556 } else if (panic_on_nmi)
557 panic("NMI indicates hardware failure");
559 #endif /* POWERFAIL_NMI */
560 #endif /* NISA > 0 */
562 case T_OFLOW: /* integer overflow fault */
567 case T_BOUND: /* bounds check fault */
575 * The kernel may have switched out the FP unit's
576 * state, causing the user process to take a fault
577 * when it tries to use the FP unit. Restore the
583 if (!pmath_emulate) {
585 ucode = FPE_FPU_NP_TRAP;
588 i = (*pmath_emulate)(&frame);
590 if (!(frame.tf_eflags & PSL_T))
592 frame.tf_eflags &= ~PSL_T;
595 /* else ucode = emulator_only_knows() XXX */
598 case T_FPOPFLT: /* FPU operand fetch fault */
603 case T_XMMFLT: /* SIMD floating-point exception */
613 case T_PAGEFLT: /* page fault */
614 (void) trap_pfault(&frame, FALSE, eva);
620 * The kernel may be using npx for copying or other
628 case T_PROTFLT: /* general protection fault */
629 case T_SEGNPFLT: /* segment not present fault */
631 * Invalid segment selectors and out of bounds
632 * %eip's and %esp's can be set up in user mode.
633 * This causes a fault in kernel mode when the
634 * kernel tries to return to user mode. We want
635 * to get this fault so that we can fix the
636 * problem here and not have to check all the
637 * selectors and pointers when the user changes
640 #define MAYBE_DORETI_FAULT(where, whereto) \
642 if (frame.tf_eip == (int)where) { \
643 frame.tf_eip = (int)whereto; \
648 * Since we don't save %gs across an interrupt
649 * frame this check must occur outside the intr
650 * nesting level check.
652 if (frame.tf_eip == (int)cpu_switch_load_gs) {
653 td->td_pcb->pcb_gs = 0;
657 if (mycpu->gd_intr_nesting_level == 0) {
659 * Invalid %fs's and %gs's can be created using
660 * procfs or PT_SETREGS or by invalidating the
661 * underlying LDT entry. This causes a fault
662 * in kernel mode when the kernel attempts to
663 * switch contexts. Lose the bad context
664 * (XXX) so that we can continue, and generate
667 MAYBE_DORETI_FAULT(doreti_iret,
669 MAYBE_DORETI_FAULT(doreti_popl_ds,
670 doreti_popl_ds_fault);
671 MAYBE_DORETI_FAULT(doreti_popl_es,
672 doreti_popl_es_fault);
673 MAYBE_DORETI_FAULT(doreti_popl_fs,
674 doreti_popl_fs_fault);
675 if (td->td_pcb->pcb_onfault) {
677 (register_t)td->td_pcb->pcb_onfault;
685 * PSL_NT can be set in user mode and isn't cleared
686 * automatically when the kernel is entered. This
687 * causes a TSS fault when the kernel attempts to
688 * `iret' because the TSS link is uninitialized. We
689 * want to get this fault so that we can fix the
690 * problem here and not every time the kernel is
693 if (frame.tf_eflags & PSL_NT) {
694 frame.tf_eflags &= ~PSL_NT;
699 case T_TRCTRAP: /* trace trap */
700 if (frame.tf_eip == (int)IDTVEC(syscall)) {
702 * We've just entered system mode via the
703 * syscall lcall. Continue single stepping
704 * silently until the syscall handler has
709 if (frame.tf_eip == (int)IDTVEC(syscall) + 1) {
711 * The syscall handler has now saved the
712 * flags. Stop single stepping it.
714 frame.tf_eflags &= ~PSL_T;
718 * Ignore debug register trace traps due to
719 * accesses in the user's address space, which
720 * can happen under several conditions such as
721 * if a user sets a watchpoint on a buffer and
722 * then passes that buffer to a system call.
723 * We still want to get TRCTRAPS for addresses
724 * in kernel space because that is useful when
725 * debugging the kernel.
727 if (user_dbreg_trap()) {
729 * Reset breakpoint bits because the
732 load_dr6(rdr6() & 0xfffffff0);
736 * Fall through (TRCTRAP kernel mode, kernel address)
740 * If DDB is enabled, let it handle the debugger trap.
741 * Otherwise, debugger traps "can't happen".
744 if (kdb_trap (type, 0, &frame))
753 # define TIMER_FREQ 1193182
757 static unsigned lastalert = 0;
759 if(time_second - lastalert > 10)
761 log(LOG_WARNING, "NMI: power fail\n");
762 sysbeep(TIMER_FREQ/880, hz);
763 lastalert = time_second;
768 #else /* !POWERFAIL_NMI */
769 /* machine/parity/power fail/"kitchen sink" faults */
770 if (isa_nmi(code) == 0) {
773 * NMI can be hooked up to a pushbutton
777 printf ("NMI ... going to debugger\n");
778 kdb_trap (type, 0, &frame);
782 } else if (panic_on_nmi == 0)
785 #endif /* POWERFAIL_NMI */
786 #endif /* NISA > 0 */
789 trap_fatal(&frame, eva);
793 /* Translate fault for emulators (e.g. Linux) */
794 if (*p->p_sysent->sv_transtrap)
795 i = (*p->p_sysent->sv_transtrap)(i, type);
797 trapsignal(p, i, ucode);
800 if (type <= MAX_TRAP_MSG) {
801 uprintf("fatal process exception: %s",
803 if ((type == T_PAGEFLT) || (type == T_PROTFLT))
804 uprintf(", fault VA = 0x%lx", (u_long)eva);
811 if (ISPL(frame.tf_cs) == SEL_UPL)
812 KASSERT(td->td_mpcount == 1, ("badmpcount trap from %p", (void *)frame.tf_eip));
814 userret(lp, &frame, sticks);
818 KKASSERT(td->td_mpcount > 0);
825 * This version doesn't allow a page fault to user space while
826 * in the kernel. The rest of the kernel needs to be made "safe"
827 * before this can be used. I think the only things remaining
828 * to be made safe are the iBCS2 code and the process tracing/
832 trap_pfault(frame, usermode, eva)
833 struct trapframe *frame;
838 struct vmspace *vm = NULL;
842 thread_t td = curthread;
843 struct proc *p = td->td_proc; /* may be NULL */
845 if (frame->tf_err & PGEX_W)
846 ftype = VM_PROT_WRITE;
848 ftype = VM_PROT_READ;
850 va = trunc_page(eva);
851 if (va < VM_MIN_KERNEL_ADDRESS) {
856 (!usermode && va < VM_MAXUSER_ADDRESS &&
857 (td->td_gd->gd_intr_nesting_level != 0 ||
858 td->td_pcb->pcb_onfault == NULL))) {
859 trap_fatal(frame, eva);
864 * This is a fault on non-kernel virtual memory.
865 * vm is initialized above to NULL. If curproc is NULL
866 * or curproc->p_vmspace is NULL the fault is fatal.
875 * Keep swapout from messing with us during this
881 * Grow the stack if necessary
883 /* grow_stack returns false only if va falls into
884 * a growable stack region and the stack growth
885 * fails. It returns true if va was not within
886 * a growable stack region, or if the stack
889 if (!grow_stack (p, va)) {
895 /* Fault in the user page: */
896 rv = vm_fault(map, va, ftype,
897 (ftype & VM_PROT_WRITE) ? VM_FAULT_DIRTY
903 * Don't allow user-mode faults in kernel address space.
909 * Since we know that kernel virtual address addresses
910 * always have pte pages mapped, we just have to fault
913 rv = vm_fault(kernel_map, va, ftype, VM_FAULT_NORMAL);
916 if (rv == KERN_SUCCESS)
920 if (mtd->td_gd->gd_intr_nesting_level == 0 &&
921 td->td_pcb->pcb_onfault) {
922 frame->tf_eip = (register_t)td->td_pcb->pcb_onfault;
925 trap_fatal(frame, eva);
929 /* kludge to pass faulting virtual address to sendsig */
932 return((rv == KERN_PROTECTION_FAILURE) ? SIGBUS : SIGSEGV);
937 trap_pfault(frame, usermode, eva)
938 struct trapframe *frame;
943 struct vmspace *vm = NULL;
947 thread_t td = curthread;
948 struct proc *p = td->td_proc;
950 va = trunc_page(eva);
951 if (va >= KERNBASE) {
953 * Don't allow user-mode faults in kernel address space.
954 * An exception: if the faulting address is the invalid
955 * instruction entry in the IDT, then the Intel Pentium
956 * F00F bug workaround was triggered, and we need to
957 * treat it is as an illegal instruction, and not a page
960 #if defined(I586_CPU) && !defined(NO_F00F_HACK)
961 if ((eva == (unsigned int)&idt[6]) && has_f00f_bug) {
962 frame->tf_trapno = T_PRIVINFLT;
972 * This is a fault on non-kernel virtual memory.
973 * vm is initialized above to NULL. If curproc is NULL
974 * or curproc->p_vmspace is NULL the fault is fatal.
985 if (frame->tf_err & PGEX_W)
986 ftype = VM_PROT_WRITE;
988 ftype = VM_PROT_READ;
990 if (map != kernel_map) {
992 * Keep swapout from messing with us during this
998 * Grow the stack if necessary
1000 /* grow_stack returns false only if va falls into
1001 * a growable stack region and the stack growth
1002 * fails. It returns true if va was not within
1003 * a growable stack region, or if the stack
1006 if (!grow_stack (p, va)) {
1012 /* Fault in the user page: */
1013 rv = vm_fault(map, va, ftype,
1014 (ftype & VM_PROT_WRITE) ? VM_FAULT_DIRTY
1020 * Don't have to worry about process locking or stacks in the kernel.
1022 rv = vm_fault(map, va, ftype, VM_FAULT_NORMAL);
1025 if (rv == KERN_SUCCESS)
1029 if (td->td_gd->gd_intr_nesting_level == 0 &&
1030 td->td_pcb->pcb_onfault) {
1031 frame->tf_eip = (register_t)td->td_pcb->pcb_onfault;
1034 trap_fatal(frame, eva);
1038 /* kludge to pass faulting virtual address to sendsig */
1039 frame->tf_err = eva;
1041 return((rv == KERN_PROTECTION_FAILURE) ? SIGBUS : SIGSEGV);
1045 trap_fatal(frame, eva)
1046 struct trapframe *frame;
1049 int code, type, ss, esp;
1050 struct soft_segment_descriptor softseg;
1052 code = frame->tf_err;
1053 type = frame->tf_trapno;
1054 sdtossd(&gdt[mycpu->gd_cpuid * NGDT + IDXSEL(frame->tf_cs & 0xffff)].sd, &softseg);
1056 if (type <= MAX_TRAP_MSG)
1057 printf("\n\nFatal trap %d: %s while in %s mode\n",
1058 type, trap_msg[type],
1059 frame->tf_eflags & PSL_VM ? "vm86" :
1060 ISPL(frame->tf_cs) == SEL_UPL ? "user" : "kernel");
1062 /* three separate prints in case of a trap on an unmapped page */
1063 printf("mp_lock = %08x; ", mp_lock);
1064 printf("cpuid = %d; ", mycpu->gd_cpuid);
1065 printf("lapic.id = %08x\n", lapic.id);
1067 if (type == T_PAGEFLT) {
1068 printf("fault virtual address = 0x%x\n", eva);
1069 printf("fault code = %s %s, %s\n",
1070 code & PGEX_U ? "user" : "supervisor",
1071 code & PGEX_W ? "write" : "read",
1072 code & PGEX_P ? "protection violation" : "page not present");
1074 printf("instruction pointer = 0x%x:0x%x\n",
1075 frame->tf_cs & 0xffff, frame->tf_eip);
1076 if ((ISPL(frame->tf_cs) == SEL_UPL) || (frame->tf_eflags & PSL_VM)) {
1077 ss = frame->tf_ss & 0xffff;
1078 esp = frame->tf_esp;
1080 ss = GSEL(GDATA_SEL, SEL_KPL);
1081 esp = (int)&frame->tf_esp;
1083 printf("stack pointer = 0x%x:0x%x\n", ss, esp);
1084 printf("frame pointer = 0x%x:0x%x\n", ss, frame->tf_ebp);
1085 printf("code segment = base 0x%x, limit 0x%x, type 0x%x\n",
1086 softseg.ssd_base, softseg.ssd_limit, softseg.ssd_type);
1087 printf(" = DPL %d, pres %d, def32 %d, gran %d\n",
1088 softseg.ssd_dpl, softseg.ssd_p, softseg.ssd_def32,
1090 printf("processor eflags = ");
1091 if (frame->tf_eflags & PSL_T)
1092 printf("trace trap, ");
1093 if (frame->tf_eflags & PSL_I)
1094 printf("interrupt enabled, ");
1095 if (frame->tf_eflags & PSL_NT)
1096 printf("nested task, ");
1097 if (frame->tf_eflags & PSL_RF)
1099 if (frame->tf_eflags & PSL_VM)
1101 printf("IOPL = %d\n", (frame->tf_eflags & PSL_IOPL) >> 12);
1102 printf("current process = ");
1104 printf("%lu (%s)\n",
1105 (u_long)curproc->p_pid, curproc->p_comm ?
1106 curproc->p_comm : "");
1110 printf("current thread = pri %d ", curthread->td_pri);
1111 if (curthread->td_pri >= TDPRI_CRIT)
1117 * we probably SHOULD have stopped the other CPUs before now!
1118 * another CPU COULD have been touching cpl at this moment...
1120 printf(" <- SMP: XXX");
1129 if ((debugger_on_panic || db_active) && kdb_trap(type, code, frame))
1132 printf("trap number = %d\n", type);
1133 if (type <= MAX_TRAP_MSG)
1134 panic("%s", trap_msg[type]);
1136 panic("unknown/reserved trap");
1140 * Double fault handler. Called when a fault occurs while writing
1141 * a frame for a trap/exception onto the stack. This usually occurs
1142 * when the stack overflows (such is the case with infinite recursion,
1145 * XXX Note that the current PTD gets replaced by IdlePTD when the
1146 * task switch occurs. This means that the stack that was active at
1147 * the time of the double fault is not available at <kstack> unless
1148 * the machine was idle when the double fault occurred. The downside
1149 * of this is that "trace <ebp>" in ddb won't work.
1154 struct mdglobaldata *gd = mdcpu;
1156 printf("\nFatal double fault:\n");
1157 printf("eip = 0x%x\n", gd->gd_common_tss.tss_eip);
1158 printf("esp = 0x%x\n", gd->gd_common_tss.tss_esp);
1159 printf("ebp = 0x%x\n", gd->gd_common_tss.tss_ebp);
1161 /* three separate prints in case of a trap on an unmapped page */
1162 printf("mp_lock = %08x; ", mp_lock);
1163 printf("cpuid = %d; ", mycpu->gd_cpuid);
1164 printf("lapic.id = %08x\n", lapic.id);
1166 panic("double fault");
1170 * Compensate for 386 brain damage (missing URKR).
1171 * This is a little simpler than the pagefault handler in trap() because
1172 * it the page tables have already been faulted in and high addresses
1173 * are thrown out early for other reasons.
1183 va = trunc_page((vm_offset_t)addr);
1185 * XXX - MAX is END. Changed > to >= for temp. fix.
1187 if (va >= VM_MAXUSER_ADDRESS)
1195 if (!grow_stack (p, va)) {
1201 * fault the data page
1203 rv = vm_fault(&vm->vm_map, va, VM_PROT_WRITE, VM_FAULT_DIRTY);
1207 if (rv != KERN_SUCCESS)
1214 * syscall2 - MP aware system call request C handler
1216 * A system call is essentially treated as a trap except that the
1217 * MP lock is not held on entry or return. We are responsible for
1218 * obtaining the MP lock if necessary and for handling ASTs
1219 * (e.g. a task switch) prior to return.
1221 * In general, only simple access and manipulation of curproc and
1222 * the current stack is allowed without having to hold MP lock.
1225 syscall2(struct trapframe frame)
1227 struct thread *td = curthread;
1228 struct proc *p = td->td_proc;
1229 struct lwp *lp = td->td_lwp;
1231 struct sysent *callp;
1232 register_t orig_tf_eflags;
1237 union sysunion args;
1240 if (ISPL(frame.tf_cs) != SEL_UPL) {
1248 KASSERT(td->td_mpcount == 0, ("badmpcount syscall from %p", (void *)frame.tf_eip));
1251 userenter(td); /* lazy raise our priority */
1253 sticks = (int)td->td_sticks;
1255 lp->lwp_md.md_regs = &frame;
1256 params = (caddr_t)frame.tf_esp + sizeof(int);
1257 code = frame.tf_eax;
1258 orig_tf_eflags = frame.tf_eflags;
1260 if (p->p_sysent->sv_prepsyscall) {
1262 * The prep code is not MP aware.
1264 (*p->p_sysent->sv_prepsyscall)(&frame, (int *)(&args.nosys.usrmsg + 1), &code, ¶ms);
1267 * Need to check if this is a 32 bit or 64 bit syscall.
1268 * fuword is MP aware.
1270 if (code == SYS_syscall) {
1272 * Code is first argument, followed by actual args.
1274 code = fuword(params);
1275 params += sizeof(int);
1276 } else if (code == SYS___syscall) {
1278 * Like syscall, but code is a quad, so as to maintain
1279 * quad alignment for the rest of the arguments.
1281 code = fuword(params);
1282 params += sizeof(quad_t);
1286 code &= p->p_sysent->sv_mask;
1287 if (code >= p->p_sysent->sv_size)
1288 callp = &p->p_sysent->sv_table[0];
1290 callp = &p->p_sysent->sv_table[code];
1292 narg = callp->sy_narg & SYF_ARGMASK;
1295 * copyin is MP aware, but the tracing code is not
1297 if (narg && params) {
1298 error = copyin(params, (caddr_t)(&args.nosys.usrmsg + 1),
1299 narg * sizeof(register_t));
1302 if (KTRPOINT(td, KTR_SYSCALL))
1303 ktrsyscall(p->p_tracep, code, narg,
1304 (void *)(&args.nosys.usrmsg + 1));
1312 * Try to run the syscall without the MP lock if the syscall
1313 * is MP safe. We have to obtain the MP lock no matter what if
1316 if ((callp->sy_narg & SYF_MPSAFE) == 0) {
1323 if (KTRPOINT(td, KTR_SYSCALL)) {
1324 ktrsyscall(p->p_tracep, code, narg, (void *)(&args.nosys.usrmsg + 1));
1329 * For traditional syscall code edx is left untouched when 32 bit
1330 * results are returned. Since edx is loaded from fds[1] when the
1331 * system call returns we pre-set it here.
1333 lwkt_initmsg(&args.lmsg, &td->td_msgport, 0,
1334 lwkt_cmd_op(code), lwkt_cmd_op_none);
1335 args.sysmsg_copyout = NULL;
1336 args.sysmsg_fds[0] = 0;
1337 args.sysmsg_fds[1] = frame.tf_edx;
1339 STOPEVENT(p, S_SCE, narg); /* MP aware */
1341 error = (*callp->sy_call)(&args);
1344 * MP SAFE (we may or may not have the MP lock at this point)
1349 * Reinitialize proc pointer `p' as it may be different
1350 * if this is a child returning from fork syscall.
1353 lp = curthread->td_lwp;
1354 frame.tf_eax = args.sysmsg_fds[0];
1355 frame.tf_edx = args.sysmsg_fds[1];
1356 frame.tf_eflags &= ~PSL_C;
1360 * Reconstruct pc, assuming lcall $X,y is 7 bytes,
1361 * int 0x80 is 2 bytes. We saved this in tf_err.
1363 frame.tf_eip -= frame.tf_err;
1368 panic("Unexpected EASYNC return value (for now)");
1371 if (p->p_sysent->sv_errsize) {
1372 if (error >= p->p_sysent->sv_errsize)
1373 error = -1; /* XXX */
1375 error = p->p_sysent->sv_errtbl[error];
1377 frame.tf_eax = error;
1378 frame.tf_eflags |= PSL_C;
1383 * Traced syscall. trapsignal() is not MP aware.
1385 if ((orig_tf_eflags & PSL_T) && !(orig_tf_eflags & PSL_VM)) {
1386 frame.tf_eflags &= ~PSL_T;
1387 trapsignal(p, SIGTRAP, 0);
1391 * Handle reschedule and other end-of-syscall issues
1393 userret(lp, &frame, sticks);
1396 if (KTRPOINT(td, KTR_SYSRET)) {
1397 ktrsysret(p->p_tracep, code, error, args.sysmsg_result);
1402 * This works because errno is findable through the
1403 * register set. If we ever support an emulation where this
1404 * is not the case, this code will need to be revisited.
1406 STOPEVENT(p, S_SCX, code);
1411 * Release the MP lock if we had to get it
1413 KASSERT(td->td_mpcount == 1, ("badmpcount syscall from %p", (void *)frame.tf_eip));
1419 * free_sysun - Put an unused sysun on the free list.
1421 static __inline void
1422 free_sysun(struct thread *td, union sysunion *sysun)
1424 struct globaldata *gd = td->td_gd;
1426 crit_enter_quick(td);
1427 sysun->lmsg.opaque.ms_sysunnext = gd->gd_freesysun;
1428 gd->gd_freesysun = sysun;
1429 crit_exit_quick(td);
1433 * sendsys2 - MP aware system message request C handler
1436 sendsys2(struct trapframe frame)
1438 struct globaldata *gd;
1439 struct thread *td = curthread;
1440 struct proc *p = td->td_proc;
1441 struct lwp *lp = td->td_lwp;
1442 register_t orig_tf_eflags;
1443 struct sysent *callp;
1444 union sysunion *sysun = NULL;
1454 if (ISPL(frame.tf_cs) != SEL_UPL) {
1462 KASSERT(td->td_mpcount == 0, ("badmpcount syscall from %p", (void *)frame.tf_eip));
1466 * access non-atomic field from critical section. p_sticks is
1467 * updated by the clock interrupt. Also use this opportunity
1468 * to lazy-raise our LWKT priority.
1471 sticks = td->td_sticks;
1473 lp->lwp_md.md_regs = &frame;
1474 orig_tf_eflags = frame.tf_eflags;
1478 * Extract the system call message. If msgsize is zero we are
1479 * blocking on a message and/or message port. If msgsize is -1
1480 * we are testing a message for completion or a message port for
1483 * The userland system call message size includes the size of the
1484 * userland lwkt_msg plus arguments. We load it into the userland
1485 * portion of our sysunion structure then we initialize the kerneland
1492 if ((msgsize = frame.tf_edx) < sizeof(struct lwkt_msg) ||
1493 msgsize > sizeof(union sysunion) - sizeof(struct sysmsg)) {
1499 * Obtain a sysun from our per-cpu cache or allocate a new one. Use
1500 * the opaque field to store the original (user) message pointer.
1501 * A critical section is necessary to interlock against interrupts
1502 * returning system messages to the thread cache.
1505 crit_enter_quick(td);
1506 if ((sysun = gd->gd_freesysun) != NULL)
1507 gd->gd_freesysun = sysun->lmsg.opaque.ms_sysunnext;
1509 sysun = malloc(sizeof(union sysunion), M_SYSMSG, M_WAITOK);
1510 crit_exit_quick(td);
1513 * Copy the user request into the kernel copy of the user request.
1515 umsg = (void *)frame.tf_ecx;
1516 error = copyin(umsg, &sysun->nosys.usrmsg, msgsize);
1519 if ((sysun->nosys.usrmsg.umsg.ms_flags & MSGF_ASYNC)) {
1524 if (max_sysmsg > 0 && lp->lwp_nsysmsg >= max_sysmsg) {
1531 * Initialize the kernel message from the copied-in data and
1532 * pull in appropriate flags from the userland message.
1534 * ms_abort_port is usually initialized in sendmsg/domsg, but since
1535 * we are not calling those functions (yet), we have to do it manually.
1537 lwkt_initmsg(&sysun->lmsg, &td->td_msgport, 0,
1538 sysun->nosys.usrmsg.umsg.ms_cmd,
1540 sysun->lmsg.ms_abort_port = sysun->lmsg.ms_reply_port;
1541 sysun->sysmsg_copyout = NULL;
1542 sysun->lmsg.opaque.ms_umsg = umsg;
1543 sysun->lmsg.ms_flags |= sysun->nosys.usrmsg.umsg.ms_flags & MSGF_ASYNC;
1546 * Extract the system call number, lookup the system call, and
1547 * set the default return value.
1549 code = (u_int)sysun->lmsg.ms_cmd.cm_op;
1550 /* We don't handle the syscall() syscall yet */
1553 free_sysun(td, sysun);
1556 if (code >= p->p_sysent->sv_size) {
1558 free_sysun(td, sysun);
1562 callp = &p->p_sysent->sv_table[code];
1564 narg = (msgsize - sizeof(struct lwkt_msg)) / sizeof(register_t);
1567 if (KTRPOINT(td, KTR_SYSCALL)) {
1568 ktrsyscall(p->p_tracep, code, narg, (void *)(&sysun->nosys.usrmsg + 1));
1571 sysun->lmsg.u.ms_fds[0] = 0;
1572 sysun->lmsg.u.ms_fds[1] = 0;
1574 STOPEVENT(p, S_SCE, narg); /* MP aware */
1577 * Make the system call. An error code is always returned, results
1578 * are copied back via ms_result32 or ms_result64. YYY temporary
1579 * stage copy p_retval[] into ms_result32/64
1581 * NOTE! XXX if this is a child returning from a fork curproc
1582 * might be different. YYY huh? a child returning from a fork
1583 * should never 'return' from this call, it should go right to the
1584 * fork_trampoline function.
1586 error = (*callp->sy_call)(sysun);
1587 gd = td->td_gd; /* RELOAD, might have switched cpus */
1591 * If a synchronous return copy p_retval to ms_result64 and return
1592 * the sysmsg to the free pool.
1594 * YYY Don't writeback message if execve() YYY
1596 sysun->nosys.usrmsg.umsg.ms_error = error;
1597 sysun->nosys.usrmsg.umsg.u.ms_fds[0] = sysun->lmsg.u.ms_fds[0];
1598 sysun->nosys.usrmsg.umsg.u.ms_fds[1] = sysun->lmsg.u.ms_fds[1];
1599 result = sysun->nosys.usrmsg.umsg.u.ms_fds[0]; /* for ktrace */
1600 if (error != 0 || code != SYS_execve) {
1602 error2 = copyout(&sysun->nosys.usrmsg.umsg.ms_copyout_start,
1603 &umsg->ms_copyout_start,
1608 if (error == EASYNC) {
1610 * Since only the current process ever messes with msgq,
1611 * we can safely manipulate it in parallel with the async
1614 TAILQ_INSERT_TAIL(&lp->lwp_sysmsgq, &sysun->sysmsg, msgq);
1616 error = (int)&sysun->sysmsg;
1619 free_sysun(td, sysun);
1622 frame.tf_eax = (register_t)error;
1625 * Traced syscall. trapsignal() is not MP aware.
1627 if ((orig_tf_eflags & PSL_T) && !(orig_tf_eflags & PSL_VM)) {
1628 frame.tf_eflags &= ~PSL_T;
1629 trapsignal(p, SIGTRAP, 0);
1633 * Handle reschedule and other end-of-syscall issues
1635 userret(lp, &frame, sticks);
1638 if (KTRPOINT(td, KTR_SYSRET)) {
1639 ktrsysret(p->p_tracep, code, error, result);
1644 * This works because errno is findable through the
1645 * register set. If we ever support an emulation where this
1646 * is not the case, this code will need to be revisited.
1648 STOPEVENT(p, S_SCX, code);
1653 * Release the MP lock if we had to get it
1655 KASSERT(td->td_mpcount == 1, ("badmpcount syscall from %p", (void *)frame.tf_eip));
1661 * waitsys2 - MP aware system message wait C handler
1664 waitsys2(struct trapframe frame)
1666 struct globaldata *gd;
1667 struct thread *td = curthread;
1668 struct proc *p = td->td_proc;
1669 struct lwp *lp = td->td_lwp;
1670 union sysunion *sysun = NULL;
1672 register_t orig_tf_eflags;
1673 int error = 0, result, sticks;
1677 if (ISPL(frame.tf_cs) != SEL_UPL) {
1685 KASSERT(td->td_mpcount == 0, ("badmpcount syscall from %p",
1686 (void *)frame.tf_eip));
1691 * access non-atomic field from critical section. p_sticks is
1692 * updated by the clock interrupt. Also use this opportunity
1693 * to lazy-raise our LWKT priority.
1696 sticks = td->td_sticks;
1698 lp->lwp_md.md_regs = &frame;
1699 orig_tf_eflags = frame.tf_eflags;
1705 TAILQ_FOREACH(ptr, &lp->lwp_sysmsgq, msgq) {
1706 if ((void *)ptr == (void *)frame.tf_ecx) {
1707 sysun = (void *)sysmsg_wait(lp,
1708 (void *)frame.tf_ecx, 1);
1718 else if (frame.tf_eax) {
1719 printf("waitport/checkport only the default port is supported at the moment\n");
1724 switch(frame.tf_edx) {
1726 sysun = (void *)sysmsg_wait(lp, NULL, 0);
1729 sysun = (void *)sysmsg_wait(lp, NULL, 1);
1738 umsg = sysun->lmsg.opaque.ms_umsg;
1739 frame.tf_eax = (register_t)sysun;
1740 sysun->nosys.usrmsg.umsg.u.ms_fds[0] = sysun->lmsg.u.ms_fds[0];
1741 sysun->nosys.usrmsg.umsg.u.ms_fds[1] = sysun->lmsg.u.ms_fds[1];
1742 sysun->nosys.usrmsg.umsg.ms_error = sysun->lmsg.ms_error;
1743 error = sysun->lmsg.ms_error;
1744 result = sysun->lmsg.u.ms_fds[0]; /* for ktrace */
1745 error = copyout(&sysun->nosys.usrmsg.umsg.ms_copyout_start,
1746 &umsg->ms_copyout_start, ms_copyout_size);
1747 free_sysun(td, sysun);
1749 code = (u_int)sysun->lmsg.ms_cmd.cm_op;
1753 frame.tf_eax = error;
1755 * Traced syscall. trapsignal() is not MP aware.
1757 if ((orig_tf_eflags & PSL_T) && !(orig_tf_eflags & PSL_VM)) {
1758 frame.tf_eflags &= ~PSL_T;
1759 trapsignal(p, SIGTRAP, 0);
1763 * Handle reschedule and other end-of-syscall issues
1765 userret(lp, &frame, sticks);
1768 if (KTRPOINT(td, KTR_SYSRET)) {
1769 ktrsysret(p->p_tracep, code, error, result);
1774 * This works because errno is findable through the
1775 * register set. If we ever support an emulation where this
1776 * is not the case, this code will need to be revisited.
1778 STOPEVENT(p, S_SCX, code);
1782 KASSERT(td->td_mpcount == 1, ("badmpcount syscall from %p",
1783 (void *)frame.tf_eip));
1789 * Simplified back end of syscall(), used when returning from fork()
1790 * directly into user mode. MP lock is held on entry and should be
1791 * released on return. This code will return back into the fork
1792 * trampoline code which then runs doreti.
1795 fork_return(p, frame)
1797 struct trapframe frame;
1801 KKASSERT(p->p_nthreads == 1);
1803 lp = LIST_FIRST(&p->p_lwps);
1805 frame.tf_eax = 0; /* Child returns zero */
1806 frame.tf_eflags &= ~PSL_C; /* success */
1810 * Newly forked processes are given a kernel priority. We have to
1811 * adjust the priority to a normal user priority and fake entry
1812 * into the kernel (call userenter()) to install a passive release
1813 * function just in case userret() decides to stop the process. This
1814 * can occur when ^Z races a fork. If we do not install the passive
1815 * release function the current process designation will not be
1816 * released when the thread goes to sleep.
1818 lwkt_setpri_self(TDPRI_USER_NORM);
1819 userenter(lp->lwp_thread);
1820 userret(lp, &frame, 0);
1822 if (KTRPOINT(lp->lwp_thread, KTR_SYSRET))
1823 ktrsysret(p->p_tracep, SYS_fork, 0, 0);
1825 p->p_flag |= P_PASSIVE_ACQ;
1827 p->p_flag &= ~P_PASSIVE_ACQ;
1829 KKASSERT(lp->lwp_thread->td_mpcount == 1);